Improved Patlak Reconstruction from Low-dose Dynamic PET Using Temporal Non-local Neural Network

Patlak model is constantly considered as a critical topic in the application of dynamic positron emission tomography (PET). Traditionally, direct reconstruction methods were preferred over indirect methods, for their complete noise modelling and comprehensive information exploited from raw data. Nevertheless, the development of direct methods was inevitably constrained by the scarcity for raw dynamic PET and the requirement for computation cost. In addition, the increasing demand for low-dose PET and the irresistible motion during the dynamic scan raise more challenges for the parametric imaging research. In this work, we proposed a deep learning based framework to compute full-dose direct Patlak reconstruction in high quality from low-dose dynamic reconstructions, with the motion correction incorporated in forward/backward projection. Moreover, a temporal non-local block was designed to complement the structural information for the reconstructed images. Results for multiple dynamic PET datasets are demonstrated in this work, compared with state-of-the-art supervised/unsupervised methods.